ReAct: Synergizing Reasoning And Acting In Language Models
- Thesis: Combine reasoning and acting to utilize the synergy between the two. With this approach, we can have reasoning traces as well as actions and handle exceptions when performing actions and interacting with the environment
- Methods:
- RaAct prompting template/technique that can be used for various problem-solving tasks
- Augment context with
thought, which LLM’s reasoning over the current context. Context is nothing but a set of previous action/feedback pairs
- Contribution:
- ReAct framework, which is prompt-based technique to combine reasoning and acting
- ReAct is easy to use, generalize to any problem domain, robust and outperform baseline, and finally aligned with humans
- Takeaways: There are great synergies between reasoning and acting. CoT is a black box that suffers from hallucination because it is not grounded with external up-to-date relevant knowledge and is dependent on the model’s internal representation for reasoning. ReAct augment this with actions by using tools and external knowledge through actions and refine reasoning based on actions.
- Improvements:
- Complex tasks may not see huge improvement with ReAct as its action space requires much more examples that can’t be fit into context window. This problem is kind of resolved with increasing context window sizes of most LLMs
#nlp #llm #agents
Back to top